Owl

Owl helps you fix a broken appliance with Alexa. It reads your machine's own manual, works out what's wrong one simple check at a time, and talks you through the repair, hands-free.

Comment

Inspiration

When an appliance breaks, most people don't repair it. They replace it, because they can't find the manual, can't read it, or don't trust themselves with it.

The numbers are stark. 58% of Americans whose large appliance broke in the last five years replaced it instead of repairing it, and 26% of those had tried to repair it first and couldn't (Consumer Reports, 2024). Only 7 of 50 appliance makers make their service manuals available (U.S. PIRG). And 2.1 million tons of major appliances went to US landfill in a single year (EPA).

The answer to most of these breakdowns is already written down, in the manufacturer's own manual. Nobody can find it with their hands inside a dishwasher. So I built Owl: it reads the manual, so you don't have to guess.

What it does

Owl is an Alexa+ add-on for fixing home appliances. You tell Alexa what the machine is doing, and Owl:

  • Answers from the manufacturer's own manual: the exact sentence and the page it's printed on, never a paraphrase or a guess.
  • Remembers your home. It knows your machines and their repair history across sessions. When the dishwasher shows E24 again, it says: "This E24 is back. Last time, on August 22, we cleared the filter."
  • Diagnoses like a technician. One simple check at a time ("start a drain cycle and listen, can you hear the pump?"), narrowing down the cause before it suggests anything.
  • Talks you through the fix hands-free, with the steps on an Echo Show-style card.
  • Orders the right part, safely. It reads the full total aloud, tax included, and charges nothing until you say yes. (The store is simulated in the demo; nothing real is charged.)
  • Knows when to say no. If the fault is behind mains wiring, it won't guide you. It tells you to call a technician.

How I built it

  • An MCP server with 22 tools (read the error, check the manual, rank the likely causes, verify the diagnosis, guide the repair, find the part, check out, record the outcome). It runs on Node.js with zero dependencies, and the same server also works as a connector in other AI assistants.
  • Amazon Nova 2 Lite on Amazon Bedrock is the agent. It chooses Owl's tools itself, live, every turn. Nothing in the demo is scripted.
  • Amazon Polly gives Owl a natural voice. The speech comes back together with each reply, so the device starts talking the moment the words appear.
  • MCP App cards show the diagnosis, the repair steps and the itemised checkout on screen-bearing devices.
  • A manual library built by visiting every major-appliance page Bosch and Siemens list. 172 manuals are read by a PDF reader I wrote from scratch, so every answer can cite its page.
  • A safety gate in code, not in the prompt: no part can be sold until the diagnosis is verified against the manual.

Measured, not claimed

  • 23 / 23 live agent scenarios passed (Nova choosing tools itself, 48 turns, no rule broken)
  • 0 parts sold on a guess, across 14,237 causes on 453 machines
  • 98.4% worst-case word recall for the PDF reader, over 1,881 pages of 30 manuals
  • 1.8 ms slowest tool (p95) on the deployed server
  • 549 checks across 48 test suites, run on every push

Challenges I ran into

  • Manuals are hostile documents. Fault tables, multi-column layouts, scanned pages. I wrote my own PDF reader and checked it word for word against a reference reader until I could trust every citation.
  • Language models love to guess. A confident wrong answer here costs someone a part they didn't need, or worse. So the rule against selling on a guess lives in code, and I tested it against every cause of every machine.
  • Voice has to feel instant. A device that pauses before it speaks feels broken, so I made the speech arrive with the reply instead of after it.
  • Keeping a live agent affordable. Every Bedrock call is metered against a spending ceiling before it's made.

What I learned

An agent is only as trustworthy as what it refuses to do. The most useful thing Owl says is sometimes "I'm not going to guess with your money," or "this one needs a technician." Quoting the manual instead of paraphrasing it turned out to be the whole product.

What's next

More manufacturers, the fault-history patterns Owl already sees across a household (the dishwasher that shows E24 three times in 240 days), and publishing Owl so anyone with an Alexa+ device can use it.

Built With

Share this project:

Updates

Submission history